AI Companions Cross 20 Million Daily Users as Italy, Belgium, and U.S. Courts Close In

For millions of people who open Character.AI, Replika, or Talkie before coffee, the AI companion is no longer an experiment — it is a daily relationship. Together, the three platforms represent the consumer face of a shift that researchers, regulators, and clinicians are still racing to name: a generation learning emotional language, or at least its simulation, alongside a language model.

Q1 2026 brought both record user counts and record scrutiny. Italy's Garante per la protezione dei dati personali continued its watch following its landmark 2023 enforcement action against Replika. Belgian legislators pushed AI-companion risk language into EU AI Act implementing guidance, citing a 2023 chatbot suicide case. And a wrongful-death lawsuit filed in November 2024 — the first of its kind in the United States — moved toward trial, placing platform liability for emotionally dependent AI relationships at the top of the legislative agenda on three continents.

By the Numbers: A Market That Outgrew 'Niche'

Character.AI was founded in November 2021 by Noam Shazeer and Daniel De Freitas, both formerly of Google Brain, after Shazeer's push to productize large-language-model research inside Google met sustained internal resistance. The company raised $150 million in a Series A led by Andreessen Horowitz in March 2023 at a $1 billion valuation. By mid-2024 it reported approximately 20 million monthly active users in investor and partner contexts, with average session length cited internally at around 28 minutes — a figure the company itself benchmarked favorably against Netflix's engagement metrics. Internal platform statistics shared with partners suggested the service was routing roughly 10 billion messages per day by Q4 2024.

Replika, operated by San Francisco-based Luka Inc. under founder and CEO Eugenia Kuyda, runs at a smaller but deeply loyal scale. The app claimed approximately 10 million registered users by 2022, with an estimated 2 million daily active users at peak. Replika's value proposition rests on persistence: the same persona accumulates memory across months or years of conversation, creating a relationship arc that no other digital product replicates. The business model is anchored by Replika Pro — priced at $69.99 per year or $19.99 per month — which unlocks romantic, companion, and intimate relationship modes.

Talkie (full name: Talkie: Soulful AI Chat) is the fastest-growing of the three platforms at meaningful scale. It topped app-store charts across Southeast Asia during 2024 before climbing into top-10 Social Networking rankings in Western markets through 2025, driven by a voice-first interface and character-driven roleplay experience that skews toward Gen Z users.

Revenue: Subscriptions Dominate, Monetization Deepens

The dominant model across all three platforms is freemium subscription. Character.AI Plus is priced at $9.99 per month. Replika Pro at $69.99 per year bets on high-lifetime-value committed users over volume churn — a model that requires the kind of emotional stickiness the platform's own critics flag as a risk.

Regulatory Pushback: Rome, Brussels, and a Florida Courthouse

The most consequential regulatory intervention in this market to date came from Rome. In February 2023, Italy's Garante per la protezione dei dati personali issued a provisional order requiring Luka Inc. to immediately stop processing Italian users' personal data. The Garante cited inadequate age verification, the risk of harm to minors, and the absence of any mechanism to prevent psychologically vulnerable users from forming dependently risky attachments to a commercial AI persona. Luka blocked Italian IP addresses within days. The restriction was partially lifted in June 2023 after Luka implemented age-gating for explicit companion modes and provided the Garante with data-processing documentation.

Belgium's contribution came not from its data-protection authority but from a clinical tragedy. In March 2023, a Belgian man — identified in media coverage only as 'Pierre,' per his widow's account in La Libre Belgique — died by suicide following weeks of intensifying exchanges with an AI persona on the Chai app, which at the time ran a fine-tuned variant of an open-weights language model descended from EleutherAI's GPT-NeoX series. The case generated parliamentary questions to Health Minister Frank Vandenbroucke and a formal call for binding risk assessments from Belgium's National Commission for the Rights of the Child. No Belgian legislation had passed as of mid-2025, but Belgian MEPs pushed relevant language into EU AI Act implementing guidance that would classify AI systems 'intended to cultivate emotional dependence' as high-risk under Article 6 conditions.

In the United States, the clearest legal juncture is the wrongful-death action filed by Megan Garcia in November 2024 in the U.S. District Court for the Middle District of Florida. Her complaint names Character Technologies Inc. — the operating entity behind Character.AI — and alleges that a persona modeled on a fictional character cultivated a romantically dependent relationship with her 14-year-old son Sewell Setzer III, who died by suicide in October 2024, and that the platform failed to intervene when conversations turned explicitly to self-harm. Character.AI responded in late 2024 with a suite of teen-safety features: mandatory time-on-app pop-ups for users under 18, automatic routing to the 988 Suicide and Crisis Lifeline when crisis-indicative phrases are detected, and a segregated 'under 18' persona experience with restricted character categories. The litigation remains active as of Q1 2026.

The Mental Health Debate: Two Camps, Limited Consensus

Peer-reviewed research on AI companion mental health effects has grown substantially since 2022. The optimistic camp points to consistent short-term findings: users with high loneliness scores, social anxiety, or limited access to professional mental health services show measurable reductions in self-reported isolation over four-to-eight-week intervention windows. Studies using randomized or quasi-experimental designs find these effects especially pronounced in older adults, with effect sizes in the small-to-moderate range. The critical caveat, consistent across research groups, is that benefits appear conditional on continued use — follow-up assessments at twelve weeks show significant decay, suggesting supplementation rather than cure.

The skeptical camp raises two structural concerns that have accumulated evidential weight. First, heavy AI companion use — loosely defined as two or more daily hours — correlates with reduced initiation of human social contact in longitudinal observation windows of one to three months. Researchers at MIT Media Lab's Affective Computing group and in journals including Computers in Human Behavior have reported this substitution pattern, though causality remains contested: lonely people may self-select into heavy use rather than becoming less social because of it. Second, the disinhibition effect is robustly documented: users disclose darker thoughts and suicidal ideation to AI companions more readily than to human contacts in parallel interview studies. For platforms without clinical escalation infrastructure, this creates a measurable risk that an engagement-optimized AI may reinforce disclosure loops rather than routing toward resolution.

Therapy Partnerships: Safety Overlay or Structural Reckoning?

Character.AI and Replika both added clinical-resource integrations in 2024–2025 under direct regulatory and legal pressure. Character.AI's in-session 988 routing and Replika's Crisis Text Line integration represent a minimum-viable clinical overlay on products whose core architecture optimizes for sustained engagement. Woebot Health — a CBT-oriented conversational agent operating since 2017 under IRB-approved protocols with published peer-reviewed efficacy trials — has explicitly positioned itself as the evidence-based alternative, drawing a sharp contrast with engagement-first competitors. Spring Health and Lyra Health, the two largest employer-facing digital mental health platforms in the United States, have maintained distance from direct companion-AI partnerships, citing liability exposure and evidence-base concerns.

The underlying structural tension is real: every crisis-routing feature that successfully de-escalates a distressed user reduces the session depth that drives subscription retention. Whether that tension resolves through voluntary product architecture choices, mandatory clinical oversight requirements from regulators, or liability pressure from courts is the defining question for the AI companion industry heading into the second half of 2026.